Implicit in much research in deep learning is the assumption that if a network achieves low loss, then it must have learned a good representation of its training data. Recent work by Kumar et al. (2025, arXiv:2505.11581) questions this assumption, arguing that networks trained with conventional gradient descent achieve low loss despite learning fractured entangled representations (FER), leaving related concepts disconnected and unrelated concepts entangled inappropriately. Kumar et al. further show that dramatically better representations are possible: a network evolved through an open-ended, human-in-the-loop process learned unified factored representations (UFR) that capture the regularities of the domain far better than a counterpart with nearly identical outputs trained with conventional gradient descent. Inspired by UFR, this paper introduces unified factored organizers (UFOs), networks trained to model regularities in the representations of other networks. A network can then be trained both to achieve low loss on its task objective and to align its representations with the model learned by the UFO. Networks trained with UFOs are shown to learn qualitatively superior representations in an image-regression domain where representations are easily interpretable. To provide quantitative evidence for the impact of UFOs on representation learning, a transformer is trained with and without UFO regularization to grok a modular arithmetic problem set. When combined with weight decay, UFOs can reduce the number of transformer training steps needed to grok by more than 10x relative to weight decay alone. Without weight decay, UFOs can reduce the number of transformer training steps needed to grok by more than 100x relative to an unregularized baseline. If UFOs are real, they suggest a new paradigm where loss functions themselves are trained to learn an effective model of good representation.
Prompt injections have emerged as one of the top security risks faced by AI agents today. Early work in prompt injection attacks looked at simple tool-call agents and static attack payloads, and were saturated since mid-2025. As of Fall 2026, human red-teaming may be nearing saturation, especially on the Gemini family of models. In this work, we argue that despite this apparent plateau, prompt injection attacks are still possible and continued effort into prompt injection defense is needed. We design a new synthetic evaluation dataset which uses random non-malicious actions as adversarial goals. Such dataset design helps us to evaluate prompt injection robustness as a model ability to respect instruction hierarchy, and disentangle this ability from model knowledge of security practices. We show that an agentic attacker of increased capabilities can reliably trick model to follow instructions in tool output in our evaluation, even for models previously considered robust on public prompt injection benchmarks such as GraySwan Arena. This shows that current frontier defenses may require proactive, automated red-teaming to remain resilient to increasingly capable adversaries.
Fleets of driving vision-language-action (VLA) models are developed one domain at a time: each operator finetunes a shared base model on private data, producing specialized experts that are costly to consolidate into a single model. Weight merging loses accuracy as heterogeneous domains accumulate, and dense mixture-of-experts routing is prohibitively expensive to combine them. We present DeltaForge, to our knowledge the first post-hoc mixture-of-experts VLA for driving, which consolidates N independently finetuned per-domain models into a single multi-domain model from their weights and a small proxy set alone. When different operators finetune the same base model, their weight updates end up sharing a lot of structure. We exploit this by decomposing all the updates together using singular value decomposition (SVD), splitting each one into a shared part (the same across all domains) and a tiny domain-specific part. The router then only needs to work with those tiny domain-specific pieces, not the full expert weights, which is what keeps the approach efficient. Then two cooperating clip-level blind routers, one over the VLM reasoning experts and one over the flow-matching dynamics experts, are trained from soft labels on the proxy dataset. Across six diverse domains spanning urban, off-road, and highway driving on cars, trucks, and all-terrain vehicles (ATVs), DeltaForge approaches the accuracy of the per-domain experts and surpasses every merging, routing, centralized, and federated baseline at close to single-model parameter and compute cost. Adding a new domain requires only projecting its weight update onto the shared basis with no end-to-end retraining needed. The resulting model also generalizes to domain combinations it was never explicitly trained on.
How many observations are needed to estimate a common rotation in two-dimensional overcomplete independent component analysis? We study equally spaced mixing columns with known source count, common source law, and column norm. For a fixed non-Gaussian fourth-order Gram-Charlier source law, the rotation is identifiable modulo the column spacing. Yet the sample size for fixed relative angular accuracy grows exponentially with the source count. We characterize this sample size, up to absolute constants, by the reciprocal of the squared Hellinger distance between two rotations separated by half the column spacing. An explicit radially weighted phase estimator attains this scale uniformly over the rotation family, using the observation radius to prevent cancellation in the mean of an angular harmonic. We also determine the exact exponential decay rate of single-observation total variation for the same pair. This gives bounds on the sample size exponent; its exact value remains open. Finally, a noisy binary-source example shows how an exact zero line of the observed characteristic function can have an exponentially small normal derivative.
Large language models have improved Verilog generation, yet functional correctness remains difficult to optimize because simulation commonly provides a coarse module-level pass/fail outcome. Such feedback discards useful evidence from partially correct RTL modules and cannot determine whether a source statement is functionally responsible for an observed output behavior. We present a new method named Causal Signal Credit Assignment (CSCA), which is a counterfactual preference-optimization framework for Verilog generation. CSCA first performs continuous signal-level behavioral verification on sampled candidate modules. It then locates output-driving assignment spans and applies five interventions while holding the remaining module and testbench fixed. The resulting changes in signal-level errors are combined with behavioral agreement to form a continuous statement credit, which is mapped to response-token weights for strict weighted direct preference optimization. This enables CSCA to learn from locally supported statements without uniformly reinforcing an otherwise incorrect response. Built on Icarus Verilog and a CodeV-derived corpus, CSCA is evaluated on VerilogEval~2.0 and RTLLM~v2.0, including both code completion and specification-to-RTL tasks. With a Qwen2.5-Coder-7B-Instruct backbone, CSCA leads the evaluated 7B-scale models and remains comparable to the much larger GPT-5.4-mini on the reported Verilog-generation benchmarks.
Medication recommendation is a multi-label prediction task. However, existing prediction-set evaluation and calibration methods rely mainly on exact-overlap metrics and therefore treat all missed medications as equally severe, overlooking how closely the prediction set matches them within the medication hierarchy. To address this, we introduce Conformal Severity Control (CSC), a post-hoc framework for calibrating medication prediction sets under hierarchy-aware mismatch severity. CSC uses the Anatomical Therapeutic Chemical (ATC) hierarchy to quantify mismatch severity and controls patient-level severity without retraining the predictor, supporting both average severity and upper-quantile severity for the highest-severity mismatches. Experiments on MIMIC-IV show that CSC tracks user-specified severity tolerances and shifts prediction sets away from distant mismatches toward exact or close matches.
Out-of-distribution (OOD) generalization remains a central challenge in materials property prediction, where models must extrapolate to novel chemistries beyond the training distribution. While large language models (LLMs) have demonstrated strong scientific reasoning, their direct use for quantitative prediction remains limited. To address this, we systematically evaluate multiple roles for LLM assistance, finding that only taxonomy-based grouping consistently improves OOD generalization over a strong kNN+GBT baseline. Building on this, we here propose MOTIF, a framework that leverages LLMs not as predictors, but as generators of candidate taxonomies that partition materials into chemically meaningful environments. These taxonomies are evaluated via training-only pseudo-OOD validation, and only those that improve generalization are retained. Invariant feature selection is then applied within each accepted taxonomy, followed by gated aggregation across taxonomies. We evaluate on 2D halide perovskites, where bandgap prediction must generalize across distinct organic spacer families. On this benchmark, MOTIF reduces mean OOD MAE from 0.122 to 0.111 eV. Under a frozen dev/test protocol, the deployable configuration further improves errors from 0.123 to 0.117 eV (electronic bandgap) and from 0.181 to 0.173 eV (optical bandgap). Beyond this primary task, we validate transferability across additional materials datasets, including 3D perovskite solar cell efficiencies and Matbench benchmarks with structure-based splits. Across domains, MOTIF consistently improves robustness under distribution shift, outperforming expert-defined groupings, rule-based heuristics, and clustering baselines, while remaining competitive with strong classical selection methods. These results demonstrate that LLMs are most effective not as predictors, but as tools for discovering structure in data.
Label-space expansion - predicting a larger joint space and summing back to the original labels - is a natural design for multi-label facial expression recognition, but its failure modes have not been characterised. We evaluate Condition-Aware Label Expansion (CALE) on a self-built dataset (30,240 training images, 3,460 multi-label test images), on seven backbones, and on the public RAF-DB and RAF-ML benchmarks. The central finding is that the aggregation rule, not the representation, decides whether the method works. With a fixed threshold of 0.5, the aggregate score spans 28x across five backbones at 10% of the training data (0.24 to 6.79) and collapses to a 1.9x range at full data, while threshold-free mAP stays within 1.28x over the same backbones, which localises the fault to the summation step. The mechanism is the uncalibrated aggregation rule, which adds the main-branch score to half the sum of the three non-main branches; we verify it by measuring main- and non-main-branch activations. Two plain seven-dimensional sigmoid baselines (BCE, ASL) reach mAP 0.5055 and 0.5175; ASL exceeds three of the five CALE configurations and falls within 0.03 of the best. On single-label FER the gain over a seven-dimensional baseline is nominally significant (paired t-test, p = 0.0436) but does not survive Benjamini-Hochberg correction (p_BH = 0.2055) or a Wilcoxon test (p = 0.1250). We conclude that sum-based label expansion requires threshold calibration at the deployment data scale, or a learned aggregation, before it can support multi-label decisions.
Connected physical systems fuse observations from robots, IoT sensors, and edge-AI modules into a shared account of the world, yet most evaluations score only the final answer or action and cannot tell whether a method attributed the evidence, updated its belief, handled contradictions, and justified its decision. We argue that shared reasoning in Connected Robotics should be evaluated as an evidence-to-belief trajectory, and introduce the Connected Robotics Evidence Graph (CREG) Benchmark to make that trajectory measurable. CREG comprises 350 deterministically generated scenarios across seven regimes (justified revision, belief preservation, relation association, contradiction onset and resolution, stale/delayed evidence, and source reliability), each exposing a normative posterior reference at every evidence prefix, validated by an independent log-odds checker and screened by a blinded expert audit. Nine representation conditions separate relation content from serialization, ordering, and metadata. Evaluating four open-weight instruction models and six transparent fusion baselines yields three findings. First, entity grounding accounts for a substantial part of the weak-model deficit. Second, deleting semantic relations costs 15.3 points of strict belief-state accuracy on relation-dependent scenarios (Holm-corrected p<0.001), but an information-matched structured list outperforms the full graph and shuffling the graph does not hurt: relation content matters where required, while graph serialization confers no consistent advantage. Third, belief accuracy is strongly non-monotonic along the evidence sequence, with mid-trajectory failures that endpoint-only evaluation cannot detect. All scenarios, references, prompts, metrics, and analysis code are released for offline, hardware-free reproduction.
An answer can match the reference for a complete agent history even when the reader's visible history does not determine it. We study this distinction using executable device logs. Our protocol separates reference agreement, certified evidence sufficiency, and context-appropriate behavior. Insufficiency requires two legal histories yielding byte-identical visible contexts but different answers; public chronology and query cutoffs certify sufficient prefixes. On 80 confirmation histories, Qwen3-32B matches 87 hidden references among 203 insufficient raw-prefix instances, with only seven appropriate abstentions. A fixed-input DeepSeek-V4-Flash replication matches 85 and abstains on 21 of the same 203 instances. Each reader makes 1,600 requests. Compact CSV's estimated appropriateness effects remain inconclusive: +3.16 percentage points, clustered 95% interval [0.00,6.57], for Qwen and +2.46, [-1.84,6.94], for DeepSeek. A separate held-out judge study achieves 99.38% correctness but retains support errors and truncations; a deterministic baseline solves all constructed cases. These results support an evaluation diagnosis across two hosted reader configurations, not gains for a general memory system or natural-language evidence assessment.
Out-of-distribution (OOD) detection is essential for deploying machine learning models in high stakes scientific applications, yet methods that flag OOD inputs without retraining the model have been validated almost exclusively on image classification. We introduce MOOD (Materials Out-of-Distribution Detection), a benchmark for evaluating whether these post-hoc OOD detectors transfer to graph neural networks (GNNs) on materials property prediction. MOOD evaluates 15 post-hoc detectors and 2 supervised diagnostic probes across 4 GNN architectures, 5 MatBench tasks, and 15 physically motivated compositional and structural shifts, yielding 264 evaluation settings and over 13,000 AUROC measurements across three encoder training seeds. Our evaluation reveals three key findings. First, current post-hoc detectors perform near chance; all 15 detectors achieve an AUROC around 0.5. Second, detector failures decompose into two regimes: method bottlenecks and encoder bottlenecks. Method bottlenecks arise when supervised probes can separate in-distribution (ID) from OOD samples but unsupervised detectors fail, meaning that the embeddings contain OOD signal that current detectors do not extract. Encoder bottlenecks arise when even supervised probes fail to separate ID from OOD samples, indicating that the relevant shift is not accessible in the frozen GNN representation. Third, when OOD prediction error exceeds ID error, the AUROC of distance-based and density-based detectors correlates with the size of that gap. Collectively, these findings indicate that reliable OOD detection in scientific domains requires representation learning strategies that explicitly preserve OOD relevant signal, particularly for tasks where current GNN architectures encode no extractable signal at all.
As language models (LMs) are increasingly deployed in downstream applications, accurately estimating their likelihood of success on a given example is critical. However, confidence estimates from LMs are poor predictors of their likelihood of success. Accurate confidence estimation is linked in part to the detection of aleatoric uncertainty; we hypothesize that LMs possess representations of at least some aspects of this uncertainty in the form of representations of the inherent ambiguity of the input. Using difference-in-means on minimal pairs of ambiguous and unambiguous sentences, we derive ambiguity representations in LM activations, and find that steering with them selectively (albeit weakly) modulates LMs' judgments of the input's ambiguity. We then propose DART, a fine-tuning method that encourages updates to lie within the span of the ambiguity directions, and penalizes changes to the LM activations outside of this subspace. DART greatly improves calibration, encodes these improvements into the model parameters, and maintains model accuracy on multiple-choice and generative QA tasks.
Unsupervised person re-identification not only faces the challenge of domain gaps, but existing methods also rarely support multiple network architectures. To tackle both issues, this paper proposes a multi-architecture auto-adaptive domain adaptation method: Mamba augmented multi-architecture domain adaptation framework. Specifically, this paper constructs the multi-architecture domain adaptation capability at two levels. At the first level, we propose a composite operator(MDA operator) that encapsulates Mamba as a component for enhancing domain adaptation. At the second level, based on this operator, we propose a Mamba-enhanced multi-architecture Domain Adaptation framework(MDA framework). Comprehensive experiments on person re-identification benchmarks validate the consistency and flexibility of the proposed framework. The MDA framework supports versatile deployment strategies without being confined to a single integration point. Under a single deployment configuration, it surpasses several SOTA methods. Moreover, it demonstrates substantial performance improvements across both CNN and Transformer architectures. When embedded into a CNN backbone, the framework achieves a 10.2% mAP improvement on the adaptation task from CUHK03 to Market-1501. Integrated with a Transformer backbone, it yields a 19.8% mAP improvement on the same task.
Vision-language models (VLMs) could support medical-image interpretation, but their use requires deciding which interpretations to accept and which images to refer to a clinician. We test whether thresholding stated confidence can retain useful coverage while limiting errors among accepted answers, and how this depends on the population. Eight open-weight VLMs, including two reasoning models, assess 3,888 brain MRI slices from post-treatment glioma patients (BraTS 2024) and healthy volunteers (IXI) for tumor or gross abnormality. BraTS labels are mask-derived; IXI slices are treated as negatives. On BraTS, no threshold selects a nonempty set with at most 5% empirical error for any model, even when chosen retrospectively. This also holds for four models evaluated on a large-lesion, distant-negative subset. At a 10% error budget, Gemma-4-12B retains 24.4% of slices and every other model at most 2.5%. To investigate this limited selection ability, we separate confidence differences between predicted answers from differences within each answer. For four models with available outcome-level summaries, pooled mean confidence differs more between answers than between correct and incorrect responses within each answer. A covariance decomposition explains how this pattern can produce opposite confidence-correctness associations across populations. At a 20% error budget, accepting every "tumor" answer covers more slices than any confidence threshold for two models. These findings caution against using unvalidated stated confidence as a standalone deferral signal and support population-specific validation before deployment. Code is available at https://github.com/amir-sbg/VLM-Answer-or-Defer.
In industrial machine vision, images captured by edge devices are often downsampled and compressed to reduce bandwidth, introducing resolution loss and compression artifacts that severely degrade the performance of downstream tasks, such as defect detection. Existing quality enhancement methods typically apply super-resolution (SR) or compression artifact removal (CAR) independently or in a sequential manner and are largely optimized for human visual fidelity rather than machine perception. This paper proposes an end-to-end quality enhancement framework that integrates SR and CAR with a unified training strategy that optimizes both human-perception and machine-perception objectives. We investigate fixed-weight and uncertainty-based dynamic-weight loss functions to balance reconstruction quality and machine vision task performance. Extensive experiments on PCB Defect, DeepPCB, and NEU-DET datasets show that the proposed jointly trained model generally outperforms standalone and separately trained SR-CAR pipelines. In particular, under severe JPEG compression, our method improves defect detection performance from 0.461 to 0.887 mAP@50 on PCB Defect over sequential SR-CAR baselines, with only a slight reduction in PSNR. These results demonstrate that joint optimization is critical for compression-aware industrial machine vision systems.
Nonseparable costs and coupled constraints in distributed online convex optimization often arise when system performance and operational constraints depend on the joint configuration of all agents and cannot be decomposed into individual contributions. Such nonseparability leads to fully distributed decision-making with challenging partial local information, where each agent evaluates its local primal-dual update using estimated global information via a local belief of the joint action and the shared Lagrange multiplier to be exchanged with neighbors. While the belief consensus established by prior works ensures no-regret learning, exchanging beliefs comprising full-dimensional primal and dual states at every round imposes a communication overhead that scales with both the joint decision dimension and the network size. Existing event-triggered schemes typically activate communication according to state variation or consensus error. Under nonseparable coupling, however, errors of equal magnitude can produce markedly different perturbations after passing through the KKT operator. We therefore propose a randomized KKT-perturbation-aware communication mechanism, where the triggering of each agent's primal and dual communication channels depends on their impact on the KKT operator. The proposed mechanism transfers the adaptive-sampling principle of coordinate descent from allocating computation across decision blocks to allocating communication across information channels: channels with larger estimated perturbations receive greater priority for refresh. Under standard regularity assumptions, the resulting algorithm achieves $O({T}^{2/3})$ expected static regret and $O({T}^{2/3})$ expected cumulative constraint violation. Numerical experiments demonstrate improved learning and constraint-enforcement performance per transmitted bit over belief consensus baselines.
When an AI diffusion-type generator samples a frustrated spin glass, which macroscopic alternatives does it decide first, and when? We study this for bridge-diffusion models of the Path Integral Diffusion (PID) type started from a point source. We show that every isotropic PID is a time-changed Gaussian stochastic localization, and conversely every signal-to-noise schedule is realized by a PID protocol. What a continuous-time generator decides therefore does not depend on the protocol, which only sets when and at what cost. We describe the generator's state by its partial commitments: the probabilities it assigns, at each moment, to nested classes of configurations fixed by the target. The partial commitments are martingales that move only along class contrasts. Their ensemble average gives the information about each level at an exact rate set by squared magnetization contrasts. Each split is decided, approximately, like a binary Gaussian channel at signal-to-noise ratio 1.04/\lambda_P, with \lambda_P the leading eigenvalue of the parent's correlation matrix. The theory is applied to Bethe-lattice and Sherrington--Kirkpatrick (SK) glasses with up to 64 spins. It is observed generically that a trajectory first shifts its weight to one of the two mirror halves of the zero-field measure and then to the glassy classes inside it. These classes appear near the glass transition and keep their identity on cooling. Nested glassy levels have similar \lambda_P and are decided together, so a trajectory commits in one or two distinct events---the mirror choice and a compound glassy decision---which separate further with coupling and size, as the spectra predict. For a generator that takes finitely many steps, the protocol matters through the late, local stage rather than the macroscopic decisions.
Combining video generation, dynamic reconstruction and monocular SLAM requires consistent coordinates, explicit handling of missing outputs and predictable memory use. We present AcrossVAM3.0, an offline system that converts text, images or monocular RGB video into a versioned point representation. Frozen components run sequentially on one GPU; map corrections update cameras, points and trajectories together. The coordination layer adds no neural parameters. Three fresh input-mode runs completed with a maximum sampled task-process-tree GPU footprint of 17.14 GB. In a frozen 16-scene video suite, every scene yielded a 4D reconstruction and 15 yielded a non-empty SLAM map. Strict cross-window matching reduced local surface error but lost long-range track coverage on TAP-Vid-DAVIS. We therefore developed multi-hypothesis adaptive surface chaining (MHAC), which retains several candidate matches and an explicit missing state. After selection on disjoint development data, MHAC improved Average Jaccard over a position-only baseline by 0.1394 on a frozen ten-video holdout (95% bootstrap interval [0.1074,0.1689]). Across 265 RoboTAP videos, the mean paired gain was 0.0610 ([0.0510,0.0715]), with 215 wins, one tie and 49 losses. These results support resource-bounded scene construction and an external tracking gain under the tested protocol. They do not establish metric scale, semantic identity or high-fidelity novel-view rendering.
We study how to combine a human binary call (a referee's offside or onside decision) with an AI numerical measurement (VAR's estimate of an attacker's position relative to the offside line). We derive an explicit maximum-likelihood rule that determines when the referee's call should stand and when a conflicting VAR measurement should overturn it. Our model treats the true offside margin as fixed and unknown and assumes independent additive Gaussian errors in the human's perception and the AI measurement. When both errors have zero means and known standard deviations $\sigma_H$ and $\sigma_{AI}$, respectively, the reversal threshold is $\sqrt{2/\pi}\,\sigma^2_{AI}/\sigma_H$. An onside call is overturned precisely when VAR's measured offside margin exceeds this threshold---not merely because VAR places the attacker beyond the offside line. We also derive the rule when either source has a systematic bias. We prove that the maximum-likelihood estimate of the true margin exists and is unique, while the final binary decision requires only a threshold comparison, without computing that estimate numerically. We illustrate the rule through numerical examples and show how to estimate the error parameters from calibration data. The same analysis applies beyond soccer to binary classification problems in which a human call and an AI measurement satisfy the stated model. This paper is intended for educational purposes: we aim to illustrate how undergraduate students and advanced high school students familiar with basic probability and calculus can investigate research questions in AI and its applications.
Fine-tuning public pretrained language models with lightweight adapters such as LoRA has become the mainstream way to adapt them to downstream tasks. A service provider often trains its adapter on a data owner's samples, so two assets need protection: the data owner's samples and the provider's adapter. Existing secure fine-tuning protocols protect the samples but reveal to the data owner intermediate results of every training step, such as the hidden representations or logits of its inputs. We show that a single step suffices to copy the adapter: from the hidden representations of one mini-batch of 32 samples, the data owner fits an adapter that agrees with the provider's on 99.2% of test inputs, and logits alone yield about 94%. Protecting both assets therefore requires that no value that depends on the adapter be revealed throughout training. We present VeilTune, to the best of our knowledge the first protocol that fine-tunes a pretrained language model with backpropagation entirely in secure multi-party computation (MPC). VeilTune computes only what the loss depends on, replaces generic MPC operators with training-aware nonlinear operators that stay accurate on the values that training produces, and obliviously skips the steps corrupted by truncation failures. On four few-shot tasks with RoBERTa-large, VeilTune nearly matches plaintext training and is 5.3 points more accurate on average than SecP-Tuning, the state-of-the-art secure fine-tuning protocol. On SST-2, it achieves a 3.5x end-to-end speedup and a 3.6x reduction in communication over SecP-Tuning, and 26x and 17x over full fine-tuning in MPC.